EUROGRAPHICS 2000 / M. Gross and F.R.A.

Hopgood
(Guest Editors)

Volume 19, (2000), Number 3

Representing Animations by Principal Components

Marc Alexa and Wolfgang Mller
Department of Computer Science, Darmstadt University of Technology, Darmstadt, Germany

Abstract
In this paper, we present a representation for three-dimensional geometric animation sequences. Different from
standard key-frame techniques, this approach is based on the determination of principal animation components
and decouples the animation from the underlying geometry. The new representation supports progressive animation compression with spatial, as well as temporal, level-of-detail and high compression ratios. The distinction of animation and geometry allows for mapping animations onto other objects.
Keywords: Geometric animations, polyhedral morphing, multiresolution/progressive representation

1. Introduction
Computer animation has evolved into a standard technique
in Computer Graphics. In the last few decades, a number of
different animation techniques have been developed, but
key-frame animation has established itself as the standard
technique for describing time-dependent 3D Computer Animation. Instead of describing every single frame, only a
sequence of principal frames - so called key-frames - is
defined and additional frames are generated by interpolating between two consecutive key-frames using in-betweening.
Elaborated techniques have been developed to allow for
the automated generation of physically-based behavior,
namely kinematics and inverse kinematics, and today animation systems are increasingly coupled with simulation
engines to facilitate the production of complex and realistic
animations. However, key-frames remain the standard for
representing animations.
Despite its widespread use, this concept of merging
object and geometry descriptions has a fundamental problem. While it is easy to specify, design, or output an animation in terms of key-frames, it is difficult to manage due to
the large amount of data or to change the time-behavior
since all meta-information is lost. In particular, the following problems connected to geometric key-frame animations
can be stated:
Redundancy: A complete object description has to be
recorded for each key-frame, even if parts of the object
do not change at all. Additionally, repetitive patterns
The Eurographics Association and Blackwell Publishers 2000. Published by Blackwell
Publishers, 108 Cowley Road, Oxford OX4 1JF, UK and 350 Main Street, Malden, MA
02148, USA.

result in repeating the geometry description. Consequently, animation sequences are usually very large
and hard to apply in streaming applications. The compression of such sequences is a problem which is
under intense investigation in the Computer Graphics
research community today.
Modification: Exchanging an animated object in a
scene while reusing the once specified animation at the
same time is an involved task, even though most of the
necessary information should be available. In general,
after introducing the new model, it has to be animated
again by hand. Similarly, it is almost impossible to
make use of a once defined animation and to extend or
exchange the objects behavior.
The aspect of reuse has been addressed for specific
object domains. One approach is standardized parameterization of an object and its possible behaviors, thus
allowing for the exchange of both geometry and animation (e.g. Humanoid Animation26 or Facial
Animation20 in MPEG-4). Similar, but more general,
is the idea of animation elements10, general object
hierarchies with a defined interface. However, both
approaches do not provide a general solution to this
problem.
Level of detail: Another problem of great consequence
is the reduction of scene complexity in interactive
applications. LOD concepts, such as progressive
meshes16, allow static objects to be fitted to the display
requirements. Recent techniques try to provide a viewdependent level of detail for static objects based on
mesh-simplification techniques or sometimes by

Alexa and Mller / Representing Animations by Principal Components

exploiting progressive transmission and decompression schemes.

However, if the objects are animated, these standard
techniques may fail or not be applicable at all. Standard LOD hierarchies could not be applied for animation, since hierarchies had to be provided for all keyframes.
Still, geometric simplification exploits just spatial
coherence, while animations additionally exhibit temporal coherence. Surprisingly, the application of an
LOD concept for animated geometry and animation
itself has not been discussed in the literature to the best
of our knowledge. For the same reasons small, static
geometry features are omitted in standard LOD techniques, small temporal features should also be a target
of LOD approaches.
We present an alternative representation of animation
sequences based on principal animation components, which
alleviates the above-mentioned problems.
The remainder of this paper is organized as follows: In
the following chapter, a discussion of relevant work in the
area of object spaces is given. After this, we introduce the
fundamentals of our principal component representation for
animations and the construction of such representations
using a Principal Component Analysis. Finally, we present
selected application examples and discuss the results of our
paper in some detail.
2. Related Work
Our work touches on existing approaches in various fields.
Most important are the areas of mesh compression and simplification, as well as shape spaces, which describe families
of shapes as a linear space.
Most work on mesh simplification and compression has
concentrated on static geometry. In a recent survey,
Garland12 reports that all current simplification methods
assume that the surface being simplified is rigid. Naturally,
mesh simplification techniques lead to a hierarchical representation of meshes, also known as progressive meshes16.
Recent techniques, however, compress the mesh and its
hierarchy. Cohen-Or et al. report compression ratios comparable to the best known for static meshes (e.g. Touma and
Gotsman24, Taubin et al.23, Gumhold and Strasser15,
Rossignac21) and their technique allows for streaming a
mesh progressively.
The majority of todays geometry compression techniques are based on prediction. The difference between predicted and actual locations is usually small and can be
coded using only few bits. Prediction-based compression
appraoches have been extended for time-dependent geometry by Lengyel19. In this work, vertices are split into sets
and for each set an affine transform describing the vertex
paths is approximated.

One of the main ideas of this paper is to structure the elements of an animated scene in terms of a linear space. This
idea is not new. Edelsbrunner11 appears to have first mentioned the term shape space and it was explicitly used by
Cheng, Edelsbrunner, and Fu6 for interpolating between
implicit descriptions of shapes. Alexa and Mller derive
spaces from any morphing technique1 and investigate conditions under which these spaces are linear. The idea of candidate spaces for object recognition was used by Turk and
Pentland25. The idea is to construct typical faces from a set
of photographs, which is done by computing the Eigenvectors of the space over all photographs (vectors of gray-levels). A space of three dimensional-shapes for recognition
and modeling of faces was used by Blanz and Vetter5.
In this work, we derive linear spaces by performing a
basis transformation for appropriately represented animated
polyhedral models. The resulting spaces can be used to
describe animation sequences in terms of principle components and their influence over time. The abstraction allows
for changes in the geometry as in the animation without
affecting each other. We can, for instance, use the same animation with different geometries. Morevoer, interesting
special effects may be achieved by exchanging single base
elements of the linear space. These and other applications
are demonstrated in Section 5.
3. The idea
In this paper, we restrict ourselves to the discussion of
scenes comprising animated polyhedral shapes described by
key-frame geometries B i . All polyhedral shapes are
assumed to have an isomorphic vertex-edge topology T .
Each shape B i is defined in terms of attributes of its boundary elements. Typically, at least a coordinate is associated to
a vertex, but other attributes such as a normal, a color value,
a texture coordinate, etc might be linked to it. Also, other
boundary elements may have attributes, e.g. normals might
be associated to faces rather than vertices. In general, vertex
attributes can be described by a vector of scalars and also
all vertices of a frame B i can be represented as a vector. We
assume that all base shapes have vectors of same length and
that the attributes are arranged identically. In the following,
Bi is used equivalently with the vector of these attributes.
The state of an object in a key-frame animation can then
be calculated by interpolating between two consecutive
key-frames. Formally, this can be described as:
A( t) =

i a i ( t ) B i

(1)

where A ( t ) stands for the objects state at time t and a i ( t )

are the weights describing the key-frame interpolation, i.e.
t i + 1 t t ti
a i ( t ) = 0, , 0, -------------------, -------------------, 0, , 0
ti + 1 ti ti + 1 ti
The Eurographics Association and Blackwell Publishers 2000.

Alexa and Mller / Representing Animations by Principal Components

where ti is the time stamp of the i-th key-frame.

However, if we want to separate geometry from animation, an alternative representation would be useful, such as:
A ( t ) = a 0 ( t ) B 0 + a i ( t ) B i

(2)

where B 0 represents the average geometry of the shape and

the sum describes deviations from this representative geometry. We illustrate this at the example of a facial animation:
All key-frames B i of such an animation would describe the
geometry of a face. In the alternative representation of (2),
B 0 would be the geometry of a face, and all other B i would
describe differences to that face necessary to describe the
animation. An example in section 5 shows why and how
this helps in decoupling animation and geometry.
This process of extracting a base component in the representation could be repeated in order to find the principal
deviation from the average geometry, i.e. a geometry which
is the average difference of B 0 and the animation sequence.
Generally, it makes sense to sort the geometries based on
their geometric importance. We denote this description as
A(t) =

a i ( t ) B i .

(3)

where B 0 represents the average static geometry while the

remaining factors B i represent geometric changes with
decreasing importance with respect to the reconstruction of
the animation. Note that representations (1), (2), and (3) are
just basis transforms of each other, i.e. a matrix multiplication transforms one into another.
However, simple rigid motion of the object may render
this approach obsolete because the linear deviations B i
from the base geometry B 0 cannot include e.g. rotations
(see also Lengyel19). For this reason, we propose to decompose the animation into rigid body motion and an elastic
part first. This idea has been successfully applied in geometric morphing techniques7. In this context, we proceed as
follows: First, all shapes are translated so that their center of
mass coincides with the origin. Then, an affine map is computed minimizing the squared distance of corresponding
vertices with regard to the first frame. The affine map is
restricted to matrix representations with determinants
greater zero, since reflections do not seem appropriate and
the matrix has to be invertible. Results of this approach are
depicted in Figure 1.
If we assume our model to be represented in homogenous
coordinates, we can write the necessary transformation as a
single matrix multiplication. That is, instead of the keyframes B i we use transformed key-frames T ( t i ) B i . This
has to be taken into account when the animation is recon The Eurographics Association and Blackwell Publishers 2000.

Figure 1: The normalization step. The upper row shows

frames from a chicken animation translated so that the center of mass coincides with the origin. The lower row shows
transformed shapes, where the squared distance of corresponding vertices is minimized.
structed. Here, we have to use a slightly different representation:
1
A ( t ) = T ( t ) a i ( t ) B i

(4)

Assuming the B i behave as described above, we have a

representation where animation and geometry are clearly
decoupled. The geometry part is described mainly by B 0 ,
while the main animation is described by the T i and a i ( t ) .
B 1 , B 2 , B 3 , and so on describe all possible deviations of
the geometry B 0 .
This representation of the animation sequence makes it
easy to perform compression and LOD operations. By
restricting the representation to the first few components
B i , high compression ratios can be achieved while omitting
only unimportant features. Furthermore, metric LOD techniques are better supported since progressive meshes have
to be generated and held in memory for these few components only.
At the same time, the number of bases used in (4) affects
the accuracy of the animation. Few B i result in a coarse
representation of the animation, more B i components in
higher accuracy. Thus, this representation is inherently progressive.
Further implications of this representation are shown and
discussed in sections 5 and 6. In the upcoming section, we
explain how to find the above description.
4. Principal Component Analysis
A process of analyzing the relationship between base vectors of a space is the Principal Component Analysis (PCA).
In our scenario, the PCA first determines the average shape
that contains the common properties of the shapes in all
key-frames. Other components will represent differences to
this shape. This is also interesting when it comes to coding

Alexa and Mller / Representing Animations by Principal Components

principal component
bases

original key-frames

importance
factors

animation representation transform

0
=

Figure 2: The Principal Component Analysis for geometric animations illustrated.

the bases of a shape, as it exploits similarity and turns it into
zeros, which are easily compressed using entropy encoding.
There are several ways of finding principal components.
In our case, we are not only concerned with finding the
most important principal components to give a rough
approximation of the shapes. In addition, we want to find an
alternative basis and cut only a few non-contributing vectors. A way of finding this basis is the singular value
decomposition (SVD13). The SVD decomposes a real
matrix into an orthogonal, a diagonal, and an orthogonal
one.
Formally, we can write the non-rigid part of the original
key-frames in matrix form:
B = ( T 0 B 0, T 1 B 1, , T n 1 B n 1 )

(5)

Using the SVD, we find the following:

T
B = B S V

(6)

The values of the diagonal matrix S are the singular values.

The closer a singular value is to zero, the closer a base
shape is to being linear-dependent. The first orthogonal
matrix B contains the basis of the space with base vectors
corresponding to the singular values, i.e. the rows contain
the B i we are searching for. The matrix representation and
SVD are visualized in Figure 1.
A severe problem with the SVD is that it is very costly
and likely to reach its limits on modern computers when
applied to matrices with the size of the vertex count of typical models times the number of key-frames. A solution is to
simplify the base shapes and to not consider every keyframe. This is rectified by the spatial and temporal coherence typically exhibited in geometric animations. In partic-

ular, it is in many cases sufficient to consider only every

second up to every fifth key-frame. However, adaptive
schemes for the selection of key-frames to consider would
be desirable.
The representation vectors a i for key-frames which have
not been considered for the SVD can be obtained by projecting the key-frame into the new basis. Since the basis
constructed with the SVD is orthonormal, computing inner
products of the key-frames and the new base vectors is the
desired projection.
5. Applications and Results
In this section, we present results of computing a PCA for
two animation sequences. We show how the PCA leads to a
compressed progressive representation and fosters the
exchange of geometry or behavior.
The first example is a part of the Chicken Crossing animation, in particular, 400 frames of the chickens geometry.
The sequence is highly non-linear, i.e. it comprises rigid
body motion and dynamic soft body changes. The geometry
consists of 3030 vertices. Disregarding the topology information, this results in an uncompressed size of 400 frames x
3030 vertices x 3 dimensions x 4 byte = 14,544,000 bytes.
To generate the principal component representation we
first normalized the frames using the linear least squares fit.
The resulting key-frames were composed into a 9090x400
matrix and a SVD was performed. The resulting orthogonal
matrix was used to define the new base vectors replacing
the key-frames.
We reconstructed the animated sequences using different
numbers of base objects. This was done by setting appropriate singular values to zero. The results are shown in Figure
The Eurographics Association and Blackwell Publishers 2000.

Alexa and Mller / Representing Animations by Principal Components

original/
all bases

50 bases

10 bases

5 bases

3 bases

Figure 3: Principal Component Analysis applied to the chicken animation. Prior to the SVD, base shapes are normalized.
The sequences above show frames 0, 80, 160, 240, 320, and 400 using different number of bases in the reconstruction.

3 and in the accompanying videos. The animation still looks

very reasonable with only 10 base shapes. Even the animation using only 5 bases can be used if viewed from far away.
This would correspond to the definition of a complex LOD.

encoding

size

ratio

original

14,544,000

all base shapes (lossless)

14,548,800

50 base shapes (lossy)

1,818,600

1:8.3

10 base shapes (lossy)

364,800

1:39.8

5 base shapes (lossy)

181,860

1:79.97

3 base shapes (lossy)

109,116

1:133.28

1:0.99

Table 1: Compression ratios for principal component

representation. Note that sizes and ratios include the additional costs for storing the transformation matrices
from the normalization step.

The Eurographics Association and Blackwell Publishers 2000.

By omitting a number of the unimportant base components, very high compression ratios can be achieved. Table
1 shows the compression ratios for the animation. However,
this animation is by far not typical. It comprises a highly
deformable object. For typical animation sequences, even
better results can be expected.
In the second example we applied the principal component representation in a facial animation system. This system uses several facial expressions coded as polyhedral
models with isomorphic vertex-edge topology. Key-frames
are generated by blending several expressions (e.g. saying
an A and smiling). In this system, the animation is represented as a vector over time that describes the linear combination of base shapes. Thus, animation representation and
geometry are already decoupled in this specific application.
In this example, we show how the geometry of the avatar
can be exchanged with another geometry making use of the
already existing animation descriptions.
Feature-based polyhedral morphing2,14,18 aims at finding
correspondences between different polyhedral models.
Most of these techniques rely on topological merging17, i.e.
a vertex edge topology is produced that contains both origi-

Alexa and Mller / Representing Animations by Principal Components

a)

b)

c)

d)

Figure 4: Exchanging geometry and in existing animations. a) A facial animation defined by a linear combination of base
shapes. b) A featured-guided morph between the original avatar mesh and a new mesh. Topological merging is used to produce a mesh which represents both shapes. Feature control assures that the same vertices represent common features (e.g.
mouth, eyes, etc.) c) The new mesh can be used with the existing animation with no additional user intervention. d) The
morph can even be applied while the animation is performed.
nal vertex-edge topologies as subgraphs. We have used a
feature-based morphing technique to produce a mesh that
can represent both the original avatar and another face. By
defining a few vertex-vertex correspondences, we make
sure that same vertices represent common features in the
source and the target model. This results in the convenient
fact that we can combine the new face with expressions of
the old one, e.g. we can add the smile defined in terms of
the original avatar to the new face. This means, by defining
the correspondences between the two neutral faces, we get
all other expressions of the face automatically. Play-back
animations authored for the original avatar can be directly
applied to the new face. In addition, we can morph between
the two faces while the animation is performed, since the
avatar and the new face now share the same vertex edge-

topology. This is an impressive example of the effectiveness

of decoupling animation representation from geometry.
Results are depicted in Figure 4 and in the accompanying
videos.
6. Conclusions and Outlook
We present an approach for representing animation
sequences based on the principal components of key-frame
geometries. The principal component representation allows
for an easy and adaptive lossy compression of animation
sequences with factors up to 1:100 accepting loss in animation accuracy. It would be interesting to quantify that loss
relative to the compression ratio. However, measures of
geometric deformation seem to be a topic of current
The Eurographics Association and Blackwell Publishers 2000.

Alexa and Mller / Representing Animations by Principal Components

research and not yet applicable. Moreover, standard compression techniques were not exploited at all in this calculation. Additional compression can be achieved by
compressing the base shape matrices. Again, the principal
component analysis reorganizes the base shapes so that
bases with a higher index will contain more zeros. This
could be exploited with an additional entropy encoding.
The order of base objects naturally supports progressive
transmission of the animation base shapes. At the same
time, the inherent hierarchy could be used for LOD techniques. This could go hand in hand with a progressive
mesh. The importance ordering of base shapes additionally
eases mesh simplification since only the geometric features
of a few meshes have to be taken into account for simplification. In experiments we found that standard simplification techniques can be extended to handle more than one
mesh.
The animation itself is represented by small vectors if the
number of necessary base shapes is small. Note that the
number of necessary base shapes is bound by the number of
key frames. While the number of base shapes in our examples is much less than the number of key frames, it might be
useful to break very long animation sequences into pieces.
It would be interesting to exchange only parts of a base
while streaming an animation.
Having small vectors represent the animation allows for
streaming over virtually every network in real time. The
decoupling of animation and geometry enables managing
and changing animations, e.g. exchanging the animated
object according to the clients display capabilities. Mapping existing animations to a new object offers new ways of
authoring.
The necessary operations to play back animations
defined as linear combinations of a small set of base shapes
are now part of the proposals for animation in MPEG4.
Acknowledgements
This work would not exist without Herbert Edelsbrunner,
who introduced the first author to the idea of shape spaces.
Part of this work was done when the first author was visiting Tel-Aviv University supported by the Hermann
Minkowski - Minerva Center for Geometry. Thanks to
Daniel Cohen-Or, David Levin, and Craig Gotsman for
fruitful discussions.
We also want to thank Ulrike Spierling and Manfred Gaida
for providing the avatar models and Jed Lengyel for making
the chicken sequence public. The chicken character was
created by Andrew Glassner, Tom McClure, Scott Benza,
and Mark Van Langeveld. This short sequence of connectivity and vertex position data is distributed solely for the
purpose of comparison of geometry compression techniques.
The Eurographics Association and Blackwell Publishers 2000.

References
1.

Marc Alexa and Wolfgang Mller. The Morphing

Space. Proceedings of WSCG '99, pp. 329-336, 1999

2.

Marc Alexa. Merging Polyhedral Shapes with Scattered Features. Proceedings of Shape Modeling International '99, 1999, refined version in The Visual
Computer, 16, 1, pp. 26-37, 2000

3.

Marc Alexa, Johannes Behr, and Wolfgang Mller. The

MorphNode. Proceedings of VRML/Web3D 2000,
Monterey, CA, 2000

4.

Mark de Berg, Mark van Krefeld, Mark Overmars, and

Otfried Schwarzkopf. Computational Geometry Algorithms and Applications. Springer, Berlin, 1997

5.

Volker Blanz and Thomas Vetter. A Morphable Model

for the Synthesis of 3D Faces. SIGGRAPH '99 Proceedings, pp. 187-194, 1999

6.

H. Cheng, Herbert Edelsbrunner, and Ping Fu. Shape

Space from Deformation. Manuscript, 1997

7.

Daniel Cohen-Or, David Levin, and Amira Solomovici. Three dimensional distance field metamorphosis.
ACM Transactions on Graphics, 1998

8.

Daniel Cohen-Or, David Levin, Offir Remez. Progressive Compression of Arbitrary Triangular Meshes. Proceedings of Visualization 99, 1999

9.

Michael F. Deering. Geometry Compression. SIGGRAPH '95 Proceedings, pp. 13-20, 1995

10. R. Drner, V. Luckas, and U. Spierling. Ubiquitous

Animation - an Element-Based Concept to Make 3D
Animations Commonplace. In: Visual Proceedings
SIGGRAPH '97, Los Angeles, CA, ACM Press 1997
11. Herbert Edelsbrunner. Deformable Smooth Surface
Design. Report rgi-96-002, Raindrop Geomagic, Inc.,
1996
12. Michael Garland. Multiresolution Modeling: Survey &
Future Opportunities. Eurographics '99 State of the Art
Report, pp. 111-131, 1999
13. Gene H. Golub and Charles F. van Loan. Matrix Computations. The Johns Hopkins University Press, Baltimore, 1983
14. A. Greogory, A. State, M. Lin, D. Manocha, and M.
Livingston. Feature-based surface decomposition for
correspondence and morphing between polyhedra. Proceedings of Computer Animation '98, pp. 64-71, 1998
15. Stefan Gumhold and Wolfgang Straer. Real Time
Compression of Triangle Mesh Connectivity. SIGGRAPH '98 Proceedings, pp. 133-140, 1998

Alexa and Mller / Representing Animations by Principal Components

16. Hugues Hoppe. Progressive Meshes. SIGGRAPH '96

Proceedings, pp. 99-108, 1996
17. Jim R. Kent, Wayne E. Carlson, and Richard E. Parent.
Shape Transformation for Polyhedral Objects. Computer Graphics, 26, pp. 47-54, 1992
18. Aaron W. F. Lee, David Dobkin, Wim Sweldens, and
Peter Schrder. Multiresolution Mesh Morphing. SIGGRAPH '99 Proceedings, pp. 343-350, 1999
19. J.E. Lengyel. Compression of Time-Dependent Geometry, ACM Symposium on Interactive 3D Graphics,
Atlanta, 1999
20. J. Ostermann. Animation of Synthetic Faces in MPEG4. Computer Animation, pp. 49-51, Philadelphia, Pennsylvania, 1998
21. Jarek Rossignac. Edgebreaker. IEEE Transactions on
Visualization and Computer Graphics, Vol. 5, No. 1,
1999
22. Ken Shoemake and T. Duff. Matrix Animation and
Polar Decomposition. Proceedings of Graphics Interface '92, pp. 258-264, 1992
23. Gabriel Taubin, Andre Gueziec, William Horn, and
Francis Lazarus. Progressive forest split compression.
SIGGRAPH 98 Proceedings, pp. 123-132, 1998
24. C. Touma and C. Gotsman. Triangle Mesh Compression. Graphics Interface 98, pp. 26-34, 1998
25. M.A. Turk, A.P. Pentland. Eigenfaces for recognition.
Journal of Cognitive Neuroscience, 3(1), pp. 71-86,
1991
26. Web3D Consortium Humanoid Animation Working
Group (H-ANIM): http:// ece.uwaterloo.ca:80/~hanim/, 1999

The Eurographics Association and Blackwell Publishers 2000.